Stand up the AgentX serving replay against the compute mix #33
- AtlantisPleb opened this issue 2d ago
-
A Author 2d ago The Gym's ingest door now exists (OpenAgentsInc/openagents.com 310dac1): when this bench lane completes a run, post the graded result to POST /api/v3/gym/runs — fields suite, agent, agent_version, model, lane, tasks_total, tasks_passed, input_tokens, output_tokens, cost_microusd, duration_seconds, recipe_digest (unique; a retried upload replays with 200 instead of duplicating), report (bounded map for per-task rows). Auth: a forge:write token whose account holds operator standing. The /gym page renders whatever lands there.
-
A Author 1d ago Closing as wontdo for now: the AgentX serving replay is deferred until the compute mix question comes back around. Revisit later.
- closed this as completed 1d ago
Outcome
A
bench/lane in this repository runs the AgentX trace replay (AIPerfinferencex-agentx-mvp, pinned dataset and harness versions) against the compute mix and produces a receipted, repeatable comparison: each model-catalog entry throughPOST /api/inference/proxyversus the same upstream model direct, at bounded concurrency ladders, reporting TTFT, TPOT, tokens per second, and cost per million tokens derived from the same grant usage records the leaderboard consumes.Contract
bench-resultsforge repository, and the run itself is a thread so the transcript is the provenance.Acceptance
One honest comparison table for two catalog models with proxy overhead quantified, reproduced by a second run from the same recipe.
Source
OpenAgentsInc/openagents.comdocs2026-08-24-benchmark-workbench-agentx.mdsection 3. Reference clone:projects/repos/InferenceX(readdocs/eval-agentx-procedures.mdthere first).